Building data set pipelines for

deep learning strategies

See important disclosures at the end of this presentation.

Ankit Awasthi

Quantitative Portfolio Manager

The Trading Show NY, September 26th, 2018

Outline of the talk

Breaking down a generic learning problem

 

Criteria for selecting datasets

 

Improving deep learning strategies using hypothetical data

 

Future Testing Strategies

 

Breaking down a general machine learning problem

Illustrative and Representative Only

Hard problems are easy (to overfit)!

Here, for a dataset of 1,000 data points, given an expected correlation of 1%, around 75% of datasets could appear to be statistically significant, whereas for expected correlation values of 13%, it is close to 0.

 

Hard problems are easy (to overfit)!

Given expected correlation of 1%, as the size of the dataset increases from 100 data points to 100,000 data points, the chances of a random dataset appearing statistically significant goes from around 92% to close 0%

 

Select a dataset that...

Is clean and reliable

 

Has high coverage

 

Has long history

 

Is predictive over long horizons

Data Cleaning and Validation

Addressing look-ahead bias

 

 

Filling in missing data

 

 

Outlier Detection

Datasets : Coverage vs History

Source : JP Morgan

Macro data and price volume data have the highest coverage with longest history!

Tremendous Success of Deep Learning

Image

Text

Speech

Abundance of Data!

Deep learning scales with data

Source : Andrew Ng, Coursera Deep learning specialization

Data Augmentation for Deep Learning

Even more data!

Much harder in Finance!

Generating Hypothetical Data for Trading

Adding noise or transformations

 

 

Theoretical models

 

 

Appropriating High Frequency Data

Stylized Effects of Financial Time Series

Absence of Autocorrelations

 

Heavy Tails

 

Volatility Clustering

 

Leverage Effect

 

Cross Correlation vs Volatility

Reference : Empirical properties of asset returns: stylized facts and statistical issues, Rama Cont

Cross-Correlation Vs Asset Volatility

Mean of 30-day rolling correlation of different assets with US Total Stock Market (VTI)

Standard Deviation is shown as error bars.

Correlation (Mean and Variance)

Correlation increases with volatility

Correlation of 30-day rolling correlation of log returns of different assets vs US Total Stock Market against  the volatility of US Total Stock Market.

Correlation

A simple example - Autoencoders

Source : Arden Dertat, Applied Deep Learning

Experimental Setup

Daily EOD Data

Training

3000 pts

Validation

1000 pts

Testing

1000 pts

Large High Frequency Data Corpus 

( 32000 pts )

Training

(High Frequency + Daily)

Validation

Out of Sample 

Evaluation

Model complexity alone doesn't help much!

One Hidden Layer : 62 -> 3 -> 62

Two Hidden Layers : 62 -> 10 -> 3 -> 10 -> 62

Three Hidden Layers : 62 -> 20 -> 10 -> 3 -> 10 -> 20 -> 62

Four Hidden Layers : 62 -> 100 -> 20 -> 10 -> 3 -> 10 -> 20 -> 100 -> 62

However....more data does!

Architecture Used : 62 -> 100 -> 20 -> 10 -> 3 -> 10 -> 20 -> 100 -> 62

Stress / Future Testing Investment Strategies

Helps to be data-driven with more data!

Scenario Analysis

 

 

Cross-Validating Hyperparameters

 

 

Future performance under different capital market assumptions

Key Takeaways

Overfitting is one of the biggest concerns when working with financial datasets

 

Deep learning works best when you have lots of data and it's the same for finance

 

Hypothetical data can be used to develop better and more robust data driven strategies

Questions?​

contact@qplum.co

ankit@qplum.co

Important Disclaimers: This presentation is the proprietary information of qplum Inc (“qplum”) and may not be disclosed or distributed to any other person without the prior consent of qplum. This information is presented for educational purposes only and does not constitute and offer to sell or a solicitation of an offer to buy any securities.  The information does not constitute investment advice and does not constitute an investment management agreement or offering circular.  

Certain information has been provided by third-party sources, and, although believed to be reliable, has not been independently verified and its accuracy or completeness cannot be guaranteed. The information is furnished as of the date shown. No representation is made with respect to its completeness or timeliness. The information is not intended to be, nor shall it be construed as, investment advice or a recommendation of any kind. Past performance is not a guarantee of future results. Important information relating to qplum and its registration with the Securities and Exchange Commission (SEC), and the National Futures Association (NFA) is available here and here. 

Building data set pipelines for deep learning strategies

By Gaurav Chakravorty

Building data set pipelines for deep learning strategies

Over the last decade deep learning has had tremendous success in pushing state of the art in numerous domains such as computer vision, natural language processing, machine translation and speech recognition. All of these domains are characterized by large quantities of data. In order to leverage deep learning research to the fullest, many progressive asset managers are experimenting with different approaches to generate and use hypothetical data so that the models can learn what to do in scenarios that the markets haven't seen yet.

  • 1,685